• A
  • A
  • A
  • ABC
  • ABC
  • ABC
  • А
  • А
  • А
  • А
  • А
Regular version of the site

Working paper

Дивизимно-агломеративный алгоритм классификации на основе минимаксной модификации частотного подхода

The conventional problem of automatic classification (AC for brevity) is considered. The suggested approach is based on the new combinations of known methods and their modifications. At first, consecutive dichotomies of the initial set are produced, whereby a family of classifications consisting of 2, 3, …, k subsets is constructed, where k is a number certainly exceeding the assumed number of classes (the divisive stage). The dichotomy used is a new modification of the frequency method, which naturally includes elements of randomization. Second, each of the constructed classifications generates a new family of classifications by consecutively uniting the subsets that are the closest to one another (the agglomerative stage). After that, only non-coinciding classifications are left for further analysis. Finally, the process is repeated several times, with the result that most of the classifications turn out to be stochastically unstable. A stable classification with the maximal number of classes is declared to be the correct solution of the initial AC, while the absence of stable classifications is interpreted as the absence of cluster structure in the initial set.